diff --git a/sds/cli.py b/sds/cli.py
index 8540d3c..78365b5 100644
--- a/sds/cli.py
+++ b/sds/cli.py
@@ -148,7 +148,24 @@ def _bench(args: argparse.Namespace, label: str, sds_bot: str | None) -> int:
         first = factions[game % 2]
         second = factions[(game + 1) % 2]
         seats, roles = [], {}
-        if sds_bot:
+        if sds_bot and getattr(args, "self_play", False):
+            # Both seats, so that one of them wins.
+            #
+            # A fit needs the label to vary. Against Princess the bot loses
+            # every game, so every label is the same number and least squares
+            # has nothing to separate - more losses do not help, because the
+            # problem is the variance and not the count. Self-play gives an
+            # even split by construction and a corpus that spans the range.
+            #
+            # The cost is the one `plan/training.md` names: two weak bots teach
+            # each other to beat a weak bot. Weights fitted this way are a
+            # starting point to be re-measured against Princess, never the
+            # number that gets reported.
+            seats.append(Seat(first, "sds", sds_bot))
+            roles[first] = "sds-A"
+            seats.append(Seat(second, "sds", sds_bot))
+            roles[second] = "sds-B"
+        elif sds_bot:
             seats.append(Seat(first, "sds", sds_bot))
             roles[first] = "sds"
             seats.append(Seat(second))
@@ -463,6 +480,12 @@ def main(argv: list[str] | None = None) -> int:
         metavar="NAME",
         help="compare this run against a recorded baseline and write comparison.md",
     )
+    bench.add_argument(
+        "--self-play",
+        action="store_true",
+        help="seat the bot on both sides; for building a training corpus whose "
+        "labels vary, not for measuring strength",
+    )
     bench.add_argument("--notes", default="", help="one line, stored with a baseline")
     bench.add_argument(
         "--suite",